Google has released a model that never writes a sentence. Hand Gemini Decision 1 a photo, a call recording or a PDF plus a list of as many as 1,000 questions, and it comes back with a probability for each answer within 150 to 250 milliseconds.
Developers can try it from 11 November through a new Gemini Decisions API in AI Studio. Because nothing is generated, only input tokens are charged, at $0.10 per million, and a free tier covers prototyping.
Questions come in three forms. A yes/no check gives the chance that a condition is true. A choice question takes 2 to 255 candidates and scores each of them. A rubric question places the input on an ordered scale that starts at 0. All answers include a confidence between 0 and 1, which lets an app hand uncertain cases to a person or a stronger model.
The demo in Google’s announcement is a kitchen counter. One photo, three questions: is there a spill, which cleanup fits, how severe is it.
One request can mix media, for instance a customer call alongside account details, or a scanned claim form next to damage photos. Image and video detail can be dialled down to save tokens.
Google has put three sandbox demos in AI Studio for trying it, and describes the target work as sorting queries, choosing tools and checking content against rules.
A request can hold up to 256K tokens of mixed media. Google says the speed barely changes as questions pile up and prices the API at close to a fifth of what Gemini 3.5 Flash-Lite costs on its Interactions API over the same benchmark suite.
Inside Google it is already a routing and safety tool. Gemini Notebook sends each query through it first, to decide which model tier and how much thinking the question deserves; Google says everyday queries get up to 40% faster. Search checks MCP tool behaviour with it, Stitch’s rapid canvas edits by voice or tap with it, and CC from Google Labs chooses among more than 200 emoji reactions with it.
The benchmark table in the announcement is less one-sided than the summary. Decision 1 is first on seven of nine benchmarks and has the lowest calibration error: 0.0250, against 0.0494 for the startup Jev’s model, 0.0721 for OpenAI’s Decisions API model and 0.2104 for Gemini 3.5 Flash-Lite. Jev’s model takes Mind2Web Multimodal and When2Call, and costs $0.68 for the seven shared evaluations against $0.88 for Gemini. All figures are Google’s.
Third-party library support is rolling out, and Gemini in Google Cloud is listed as coming soon without a date.
